Claude Opus 4.8 strengthens agents, but costs remain a problem

On June 2, 2026, Ben’s Bites summarised the release of Claude Opus 4.8 and described it as an important step forward for work in Claude Code. According to the newsletter, the release introduces dynamic workflows: Claude writes a control script and then launches smaller sub-agents that solve parts of a more complex task in parallel. For developers, this means less manual division of work. For companies, it means agents will look more like internal automation teams than a single chatbot. The newsletter also quotes a more cautious view. According to Ben’s Bites, Dex notes that this does not prove the viability of loosely connected multi-agent systems. More reliable, in his view, are deterministic procedures built around smaller agentic loops. That is an important point for enterprise deployment. A company does not need an agent to look impressive. It needs it to repeatedly deliver the same type of output, be checkable and not get lost in a long chain of decisions. Ben’s Bites says Simon Willison called Opus 4.8 a modest but useful upgrade. The reason is said to be a greater willingness by the model to admit uncertainty and a lower tendency to overlook errors in its own code. For real work, that is more valuable than another impressive demonstration. When a model finds a weakness in its own proposal, it saves time for a senior person who would otherwise have to catch the error during review or in production. The newsletter also mentions the Datacurve DeepSWE benchmark. According to the text, it ranks Opus 4.8 below GPT-5.5 and only slightly above GPT-5.4, while pointing to higher token use. In company budget terms, better capability alone is not enough. If an agent completes a task but uses too much computing resource, it may be more expensive than human work or simpler automation. In the same issue, Ben’s Bites also covers news from Nvidia and Microsoft: RTX Spark is supposed to offer one petaflop of performance for local AI on Windows, up to 128 GB of unified memory, CUDA and RTX support, and local models of up to 120 billion parameters. This shows a second direction in the market. Part of AI work may move back from the cloud to devices if companies want lower latency, better data control and less volatile costs. For Czech companies, the practical question for pilots is clear: do not measure only output quality, but also the number of steps, consumption, required review and price per completed task. The model that looks smartest in a demonstration may not be the best in a company process. In 2026, the difference between it can do it and it pays off is becoming much sharper. Terms: A sub-agent is a smaller agent launched by the main system for part of a task. A token is a basic unit of text often used to calculate model cost. Local AI means a model running on a company or user device, not only in the cloud.

Claude Opus 4.8 strengthens agents, but costs remain a problem

The full article is available to users registered at Hard Skills